-
Notifications
You must be signed in to change notification settings - Fork 67
/
Copy pathllava.py
executable file
·239 lines (205 loc) · 9.12 KB
/
llava.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
# https://github.com/Woolverine94/biniou
# llava.py
import gradio as gr
import os
from llama_cpp import Llama
from llama_cpp.llama_chat_format import Llava15ChatHandler
from PIL import Image
from huggingface_hub import snapshot_download, hf_hub_download
from ressources.common import *
from ressources.tools import biniouUIControl
import multiprocessing
device_label_llava, model_arch = detect_device()
device_llava = torch.device(device_label_llava)
# Gestion des modèles
model_path_llava = "./models/llava/"
os.makedirs(model_path_llava, exist_ok=True)
model_list_llava = {}
# for filename in os.listdir(model_path_llava):
# f = os.path.join(model_path_llava, filename)
# if os.path.isfile(f) and filename.endswith('.gguf') :
# final_f = {f:(f, "")}
# model_list_llava.update(final_f)
model_list_llava_builtin = {
"xtuner/llava-llama-3-8b-v1_1-gguf":("llava-llama-3-8b-v1_1-int4.gguf", "llava-llama-3-8b-v1_1-mmproj-f16.gguf", "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system}<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n{prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>", "You are a very helpful assistant. You are accurate, concise, reliable and you never refuse to answer."),
"xtuner/llava-phi-3-mini-gguf":("llava-phi-3-mini-int4.gguf", "llava-phi-3-mini-mmproj-f16.gguf", "{prompt}",""),
"mys/ggml_bakllava-1":("ggml-model-q5_k.gguf", "mmproj-model-f16.gguf", "{prompt}", ""),
"cmp-nct/llava-1.6-gguf":("ggml-mistral-7b-q_5_k.gguf", "mmproj-mistral7b-f16.gguf", "{prompt}", ""),
"mys/ggml_llava-v1.5-7b":("ggml-model-q5_k.gguf", "mmproj-model-f16.gguf", "{prompt}", ""),
"mys/ggml_llava-v1.5-13b":("ggml-model-q5_k.gguf", "mmproj-model-f16.gguf", "{prompt}", ""),
"bee-kake/nanollava-1.5-gguf":("nanollava-text-model-f16.gguf", "nanollava-mmproj-f16.gguf", "{prompt}", ""),
}
model_list_llava.update(model_list_llava_builtin)
def download_model(modelid_llava):
if modelid_llava[0:9] != "./models/":
hf_hub_path_llava = hf_hub_download(
repo_id=modelid_llava,
filename=model_list_llava[modelid_llava][0],
repo_type="model",
cache_dir=model_path_llava,
resume_download=True,
local_files_only=True if offline_test() else None
)
modelid_llava = hf_hub_path_llava
return modelid_llava
def download_mmproj(modelid_mmproj):
if modelid_mmproj[0:9] != "./models/":
hf_hub_path_llava = hf_hub_download(
repo_id=modelid_mmproj,
filename=model_list_llava[modelid_mmproj][1],
repo_type="model",
cache_dir=model_path_llava,
resume_download=True,
local_files_only=True if offline_test() else None
)
modelid_mmproj = hf_hub_path_llava
return modelid_mmproj
@metrics_decoration
def text_llava(
modelid_llava,
max_tokens_llava,
seed_llava,
stream_llava,
n_ctx_llava,
repeat_penalty_llava,
temperature_llava,
top_p_llava,
top_k_llava,
img_llava,
prompt_llava,
history_llava,
prompt_template_llava,
system_template_llava,
progress_txt2vid_ze=gr.Progress(track_tqdm=True)
):
print(">>>[Llava 👁️ ]: starting answer generation")
modelid_llava_origin = modelid_llava
modelid_llava = download_model(modelid_llava_origin)
modelid_mmproj_llava = download_mmproj(modelid_llava_origin)
image_url = "https://localhost:7860/file="+ img_llava
if prompt_template_llava == "" :
prompt_template_llava = "{prompt}"
prompt_full_llava = prompt_template_llava.replace("{prompt}", prompt_llava)
prompt_full_llava = prompt_full_llava.replace("{system}", system_template_llava)
prompt_full_llava = prompt_full_llava.replace("{system_message}", system_template_llava)
if history_llava != "[]" :
history_final = ""
for i in range(len(history_llava)):
history_final += history_llava[i][0]+ "\n"
history_final += history_llava[i][1]+ "\n"
prompt_final_llava = f"{history_final}\n{prompt_full_llava}"
else :
prompt_final_llava = prompt_full_llava
chat_handler_llava = Llava15ChatHandler(clip_model_path=modelid_mmproj_llava)
if (biniouUIControl.detect_llama_backend() == "cuda"):
llm = Llama(model_path=modelid_llava, seed=seed_llava, n_gpu_layers=-1, n_threads=multiprocessing.cpu_count(), n_threads_batch=multiprocessing.cpu_count(), n_ctx=n_ctx_llava, chat_handler=chat_handler_llava, logits_all=True)
else:
llm = Llama(model_path=modelid_llava, seed=seed_llava, n_ctx=n_ctx_llava, chat_handler=chat_handler_llava, logits_all=True)
if system_template_llava == "":
system_template_llava = "You are an assistant who perfectly describes images."
# {"role": "system", "content": "You are an assistant who perfectly describes images."},
messages_llava = [
{"role": "system", "content": system_template_llava},
{
"role": "user",
"content": [
{"type": "image_url", "image_url": {"url": image_url}},
{"type" : "text", "text": prompt_final_llava}
]
}
]
output_llava = chat_handler_llava(
llama=llm,
messages=messages_llava,
max_tokens=max_tokens_llava,
repeat_penalty=repeat_penalty_llava,
temperature=temperature_llava,
top_p=top_p_llava,
top_k=top_k_llava,
echo=True
)
answer_llava = (output_llava["choices"][0]["message"]["content"])
last_answer_llava = answer_llava.replace(f"{prompt_final_llava}", "")
filename_llava = write_seeded_file(seed_llava, history_final, prompt_llava, last_answer_llava)
history_llava.append((prompt_llava, last_answer_llava))
print(f">>>[Llava 👁️ ]: generated 1 answer")
reporting_llava = f">>>[Llava 👁️ ]: "+\
f"Settings : Model={modelid_llava_origin} | "+\
f"Max tokens={max_tokens_llava} | "+\
f"Stream results={stream_llava} | "+\
f"n_ctx={n_ctx_llava} | "+\
f"Repeat penalty={repeat_penalty_llava} | "+\
f"Temperature={temperature_llava} | "+\
f"Top_k={top_k_llava} | "+\
f"Top_p={top_p_llava} | "+\
f"Prompt template={prompt_template_llava} | "+\
f"Prompt={prompt_llava} | "+\
f"Seed={seed_llava}"
print(reporting_llava)
metadata_writer_txt(reporting_llava, filename_llava)
del chat_handler_llava, llm, output_llava
clean_ram()
print(f">>>[Llava 👁️ ]: leaving module")
return history_llava, history_llava[-1][1], filename_llava
@metrics_decoration
def text_llava_continue(
modelid_llava,
max_tokens_llava,
seed_llava,
stream_llava,
n_ctx_llava,
repeat_penalty_llava,
temperature_llava,
top_p_llava,
top_k_llava,
img_llava,
history_llava,
):
print(">>>[Llava 👁️ ]: continuing answer generation")
modelid_llava_origin = modelid_llava
modelid_llava = download_model(modelid_llava)
if history_llava != "[]" :
history_final = ""
for i in range(len(history_llava)) :
history_final += history_llava[i][0]+ "\n"
history_final += history_llava[i][1]+ "\n"
history_final = history_final.rstrip()
if (biniouUIControl.detect_llama_backend() == "cuda"):
llm = Llama(model_path=modelid_llava, seed=seed_llava, n_gpu_layers=-1, n_threads=multiprocessing.cpu_count(), n_threads_batch=multiprocessing.cpu_count(), n_ctx=n_ctx_llava)
else:
llm = Llama(model_path=modelid_llava, seed=seed_llava, n_ctx=n_ctx_llava)
output_llava = llm.create_completion(
f"{history_final}",
max_tokens=max_tokens_llava,
stream=stream_llava,
repeat_penalty=repeat_penalty_llava,
temperature=temperature_llava,
top_p=top_p_llava,
top_k=top_k_llava,
)
answer_llava = (output_llava['choices'][0]['text'])
last_answer_llava = answer_llava.replace(f"{history_final}", "")
last_answer_llava = last_answer_llava.replace("<|im_end|>", "")
last_answer_llava = last_answer_llava.replace("<|im_start|>user", "")
last_answer_llava = last_answer_llava.replace("<|im_start|>assistant", "")
global_answer_llava = f"{history_final}{answer_llava}"
filename_llava = write_seeded_file(seed_llava, global_answer_llava)
history_llava[-1][1] += last_answer_llava
# history_llava.append((prompt_llava, last_answer_llava))
print(f">>>[Llava 👁️ ]: continued 1 answer")
reporting_llava = f">>>[Llava 👁️ ]: "+\
f"Settings : Model={modelid_llava_origin} | "+\
f"Max tokens={max_tokens_llava} | "+\
f"Stream results={stream_llava} | "+\
f"n_ctx={n_ctx_llava} | "+\
f"Repeat penalty={repeat_penalty_llava} | "+\
f"Temperature={temperature_llava} | "+\
f"Top_p={top_p_llava} | "+\
f"Top_k={top_k_llava} | "+\
f"Seed={seed_llava}"
print(reporting_llava)
metadata_writer_txt(reporting_llava, filename_llava)
del llm, output_llava
clean_ram()
print(f">>>[Llava 👁️ ]: leaving module")
return history_llava, history_llava[-1][1], filename_llava